Bridging Locomotion and Manipulation Using Reconfigurable Robotic Limbs via Reinforcement Learning

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Locomotion and manipulation are key robotic skills, often studied separately.
• A topological duality exists between legged locomotion and multi-fingered manipulation.
• Data-driven evidence for this duality and unified skill learning is lacking.

Purpose of the Study:

• To explore a unified formulation of the loco-manipulation problem using reinforcement learning (RL).
• To investigate the transferability of locomotion and manipulation skills within a single RL policy.
• To provide data-driven evidence for unified loco-manipulation and demonstrate Sim2Real transfer.

Main Methods:

• Utilized an overconstrained robotic limb design for reconfiguration into multi-legged and multi-fingered robots.
• Implemented a co-training architecture for reinforcement learning to develop a unified loco-manipulation policy.
• Employed multilayer perceptron (MLP) and graph neural network (GNN) architectures for the RL policy.

Main Results:

• Found data-driven evidence supporting skill transferability between locomotion and manipulation using a single RL policy.
• Demonstrated successful Sim2Real transfer of learned loco-manipulation skills on a physical robotic prototype.
• Validated the unified formulation through experimental results on a novel platform.

Conclusions:

• The study provides strong learning-based evidence for the transferability of loco-manipulation skills.
• A unified RL approach effectively addresses both locomotion and manipulation tasks.
• The findings open new avenues for developing more versatile and adaptable robotic systems.